zqz/platform-external-webrtc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

External denoiser based on noise estimation and moving object detection.

Browse Source 
Improved the existing external denoiser in WebRTC: the filter strength
is adaptive based on the noise level of the whole frame and the moving
object detection result. The adaptive filter effectively removes the
artifacts in previous version, such as trailing and blockiness on moving
objects.
The external denoiser is off by default for now.

BUG=

Review URL: https://codereview.webrtc.org/1822333003

Cr-Commit-Position: refs/heads/master@{#12198}
This commit is contained in:
jackychen
2016-04-01 07:46:58 -07:00
committed by Commit bot
parent 14847d15c0
commit fa0befe13b
17 changed files with 488 additions and 122 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
webrtc/modules/video_processing/BUILD.gn
View File

@ -29,6 +29,8 @@ source_set("video_processing") {
"util/denoiser_filter.h",
"util/denoiser_filter_c.cc",
"util/denoiser_filter_c.h",
"util/noise_estimation.cc",
"util/noise_estimation.h",
"util/skin_detection.cc",
"util/skin_detection.h",
"video_decimator.cc",

4
webrtc/modules/video_processing/frame_preprocessor.cc
View File

@ -22,6 +22,7 @@ VPMFramePreprocessor::VPMFramePreprocessor()
spatial_resampler_ = new VPMSimpleSpatialResampler();
ca_ = new VPMContentAnalysis(true);
vd_ = new VPMVideoDecimator();
EnableDenosing(false);
}
VPMFramePreprocessor::~VPMFramePreprocessor() {
@ -115,7 +116,8 @@ const VideoFrame* VPMFramePreprocessor::PreprocessFrame(
const VideoFrame* current_frame = &frame;
if (denoiser_) {
denoiser_->DenoiseFrame(*current_frame, &denoised_frame_);
denoiser_->DenoiseFrame(*current_frame, &denoised_frame_,
&denoised_frame_prev_, 0);
current_frame = &denoised_frame_;
}

1
webrtc/modules/video_processing/frame_preprocessor.h
View File

@ -71,6 +71,7 @@ class VPMFramePreprocessor {
VideoContentMetrics* content_metrics_;
VideoFrame denoised_frame_;
VideoFrame denoised_frame_prev_;
VideoFrame resampled_frame_;
VPMSpatialResampler* spatial_resampler_;
VPMContentAnalysis* ca_;

61
webrtc/modules/video_processing/test/denoiser_test.cc
View File

@ -21,8 +21,9 @@
namespace webrtc {
TEST_F(VideoProcessingTest, CopyMem) {
std::unique_ptr<DenoiserFilter> df_c(DenoiserFilter::Create(false));
std::unique_ptr<DenoiserFilter> df_sse_neon(DenoiserFilter::Create(true));
std::unique_ptr<DenoiserFilter> df_c(DenoiserFilter::Create(false, nullptr));
std::unique_ptr<DenoiserFilter> df_sse_neon(
DenoiserFilter::Create(true, nullptr));
uint8_t src[16 * 16], dst[16 * 16];
for (int i = 0; i < 16; ++i) {
for (int j = 0; j < 16; ++j) {
@ -48,8 +49,9 @@ TEST_F(VideoProcessingTest, CopyMem) {
}
TEST_F(VideoProcessingTest, Variance) {
std::unique_ptr<DenoiserFilter> df_c(DenoiserFilter::Create(false));
std::unique_ptr<DenoiserFilter> df_sse_neon(DenoiserFilter::Create(true));
std::unique_ptr<DenoiserFilter> df_c(DenoiserFilter::Create(false, nullptr));
std::unique_ptr<DenoiserFilter> df_sse_neon(
DenoiserFilter::Create(true, nullptr));
uint8_t src[16 * 16], dst[16 * 16];
uint32_t sum = 0, sse = 0, var;
for (int i = 0; i < 16; ++i) {
@ -71,51 +73,53 @@ TEST_F(VideoProcessingTest, Variance) {
}
TEST_F(VideoProcessingTest, MbDenoise) {
std::unique_ptr<DenoiserFilter> df_c(DenoiserFilter::Create(false));
std::unique_ptr<DenoiserFilter> df_sse_neon(DenoiserFilter::Create(true));
uint8_t running_src[16 * 16], src[16 * 16], dst[16 * 16], dst_ref[16 * 16];
std::unique_ptr<DenoiserFilter> df_c(DenoiserFilter::Create(false, nullptr));
std::unique_ptr<DenoiserFilter> df_sse_neon(
DenoiserFilter::Create(true, nullptr));
uint8_t running_src[16 * 16], src[16 * 16];
uint8_t dst[16 * 16], dst_sse_neon[16 * 16];
// Test case: |diff| <= |3 + shift_inc1|
for (int i = 0; i < 16; ++i) {
for (int j = 0; j < 16; ++j) {
running_src[i * 16 + j] = i * 11 + j;
src[i * 16 + j] = i * 11 + j + 2;
dst_ref[i * 16 + j] = running_src[i * 16 + j];
}
}
memset(dst, 0, 16 * 16);
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
EXPECT_EQ(0, memcmp(dst, dst_ref, 16 * 16));
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1, false);
memset(dst_sse_neon, 0, 16 * 16);
df_sse_neon->MbDenoise(running_src, 16, dst_sse_neon, 16, src, 16, 0, 1,
false);
EXPECT_EQ(0, memcmp(dst, dst_sse_neon, 16 * 16));
// Test case: |diff| >= |4 + shift_inc1|
for (int i = 0; i < 16; ++i) {
for (int j = 0; j < 16; ++j) {
running_src[i * 16 + j] = i * 11 + j;
src[i * 16 + j] = i * 11 + j + 5;
dst_ref[i * 16 + j] = src[i * 16 + j] - 2;
}
}
memset(dst, 0, 16 * 16);
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
EXPECT_EQ(0, memcmp(dst, dst_ref, 16 * 16));
memset(dst, 0, 16 * 16);
df_sse_neon->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
EXPECT_EQ(0, memcmp(dst, dst_ref, 16 * 16));
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1, false);
memset(dst_sse_neon, 0, 16 * 16);
df_sse_neon->MbDenoise(running_src, 16, dst_sse_neon, 16, src, 16, 0, 1,
false);
EXPECT_EQ(0, memcmp(dst, dst_sse_neon, 16 * 16));
// Test case: |diff| >= 8
for (int i = 0; i < 16; ++i) {
for (int j = 0; j < 16; ++j) {
running_src[i * 16 + j] = i * 11 + j;
src[i * 16 + j] = i * 11 + j + 8;
dst_ref[i * 16 + j] = src[i * 16 + j] - 6;
}
}
memset(dst, 0, 16 * 16);
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
EXPECT_EQ(0, memcmp(dst, dst_ref, 16 * 16));
memset(dst, 0, 16 * 16);
df_sse_neon->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
EXPECT_EQ(0, memcmp(dst, dst_ref, 16 * 16));
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1, false);
memset(dst_sse_neon, 0, 16 * 16);
df_sse_neon->MbDenoise(running_src, 16, dst_sse_neon, 16, src, 16, 0, 1,
false);
EXPECT_EQ(0, memcmp(dst, dst_sse_neon, 16 * 16));
// Test case: |diff| > 15
for (int i = 0; i < 16; ++i) {
@ -126,9 +130,10 @@ TEST_F(VideoProcessingTest, MbDenoise) {
}
memset(dst, 0, 16 * 16);
DenoiserDecision decision =
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
df_c->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1, false);
EXPECT_EQ(COPY_BLOCK, decision);
decision = df_sse_neon->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1);
decision =
df_sse_neon->MbDenoise(running_src, 16, dst, 16, src, 16, 0, 1, false);
EXPECT_EQ(COPY_BLOCK, decision);
}
@ -138,7 +143,9 @@ TEST_F(VideoProcessingTest, Denoiser) {
// Create SSE or NEON denoiser.
VideoDenoiser denoiser_sse_neon(true);
VideoFrame denoised_frame_c;
VideoFrame denoised_frame_track_c;
VideoFrame denoised_frame_sse_neon;
VideoFrame denoised_frame_track_sse_neon;
std::unique_ptr<uint8_t[]> video_buffer(new uint8_t[frame_length_]);
while (fread(video_buffer.get(), 1, frame_length_, source_file_) ==
@ -147,8 +154,10 @@ TEST_F(VideoProcessingTest, Denoiser) {
EXPECT_EQ(0, ConvertToI420(kI420, video_buffer.get(), 0, 0, width_, height_,
0, kVideoRotation_0, &video_frame_));
denoiser_c.DenoiseFrame(video_frame_, &denoised_frame_c);
denoiser_sse_neon.DenoiseFrame(video_frame_, &denoised_frame_sse_neon);
denoiser_c.DenoiseFrame(video_frame_, &denoised_frame_c,
&denoised_frame_track_c, -1);
denoiser_sse_neon.DenoiseFrame(video_frame_, &denoised_frame_sse_neon,
&denoised_frame_track_sse_neon, -1);
// Denoising results should be the same for C and SSE/NEON denoiser.
ASSERT_TRUE(test::FramesEqual(denoised_frame_c, denoised_frame_sse_neon));

13
webrtc/modules/video_processing/util/denoiser_filter.cc
View File

@ -18,13 +18,16 @@
namespace webrtc {
const int kMotionMagnitudeThreshold = 8 * 3;
const int kSumDiffThreshold = 16 * 16 * 2;
const int kSumDiffThresholdHigh = 600;
const int kSumDiffThreshold = 96;
const int kSumDiffThresholdHigh = 448;
std::unique_ptr<DenoiserFilter> DenoiserFilter::Create(
bool runtime_cpu_detection) {
bool runtime_cpu_detection,
CpuType* cpu_type) {
std::unique_ptr<DenoiserFilter> filter;
if (cpu_type != nullptr)
*cpu_type = CPU_NOT_NEON;
if (runtime_cpu_detection) {
// If we know the minimum architecture at compile time, avoid CPU detection.
#if defined(WEBRTC_ARCH_X86_FAMILY)
@ -40,9 +43,13 @@ std::unique_ptr<DenoiserFilter> DenoiserFilter::Create(
#endif
#elif defined(WEBRTC_HAS_NEON)
filter.reset(new DenoiserFilterNEON());
if (cpu_type != nullptr)
*cpu_type = CPU_NEON;
#elif defined(WEBRTC_DETECT_NEON)
if (WebRtc_GetCPUFeaturesARM() & kCPUFeatureNEON) {
filter.reset(new DenoiserFilterNEON());
if (cpu_type != nullptr)
*cpu_type = CPU_NEON;
} else {
filter.reset(new DenoiserFilterC());
}

8
webrtc/modules/video_processing/util/denoiser_filter.h
View File

@ -11,6 +11,7 @@
#ifndef WEBRTC_MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_H_
#define WEBRTC_MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_H_
#include <climits>
#include <memory>
#include "webrtc/modules/include/module_common_types.h"
@ -23,6 +24,7 @@ extern const int kSumDiffThreshold;
extern const int kSumDiffThresholdHigh;
enum DenoiserDecision { COPY_BLOCK, FILTER_BLOCK };
enum CpuType { CPU_NEON, CPU_NOT_NEON };
struct DenoiseMetrics {
uint32_t var;
uint32_t sad;
@ -32,7 +34,8 @@ struct DenoiseMetrics {
class DenoiserFilter {
public:
static std::unique_ptr<DenoiserFilter> Create(bool runtime_cpu_detection);
static std::unique_ptr<DenoiserFilter> Create(bool runtime_cpu_detection,
CpuType* cpu_type);
virtual ~DenoiserFilter() {}
@ -56,7 +59,8 @@ class DenoiserFilter {
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) = 0;
int increase_denoising,
bool denoise_always) = 0;
};
} // namespace webrtc

10
webrtc/modules/video_processing/util/denoiser_filter_c.cc
View File

@ -66,7 +66,8 @@ DenoiserDecision DenoiserFilterC::MbDenoise(uint8_t* mc_running_avg_y,
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) {
int increase_denoising,
bool denoise_always) {
int sum_diff_thresh = 0;
int sum_diff = 0;
int adj_val[3] = {3, 4, 6};
@ -136,9 +137,12 @@ DenoiserDecision DenoiserFilterC::MbDenoise(uint8_t* mc_running_avg_y,
sum_diff += col_sum[c];
}
sum_diff_thresh = kSumDiffThreshold;
if (increase_denoising)
if (denoise_always)
sum_diff_thresh = INT_MAX;
else if (increase_denoising)
sum_diff_thresh = kSumDiffThresholdHigh;
else
sum_diff_thresh = kSumDiffThreshold;
if (abs(sum_diff) > sum_diff_thresh) {
int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
// Only apply the adjustment for max delta up to 3.

3
webrtc/modules/video_processing/util/denoiser_filter_c.h
View File

@ -38,7 +38,8 @@ class DenoiserFilterC : public DenoiserFilter {
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) override;
int increase_denoising,
bool denoise_always) override;
};
} // namespace webrtc

12
webrtc/modules/video_processing/util/denoiser_filter_neon.cc
View File

@ -106,13 +106,15 @@ DenoiserDecision DenoiserFilterNEON::MbDenoise(uint8_t* mc_running_avg_y,
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) {
int increase_denoising,
bool denoise_always) {
// If motion_magnitude is small, making the denoiser more aggressive by
// increasing the adjustment for each level, level1 adjustment is
// increased, the deltas stay the same.
int shift_inc =
(increase_denoising && motion_magnitude <= kMotionMagnitudeThreshold) ? 1
: 0;
int sum_diff_thresh = 0;
const uint8x16_t v_level1_adjustment = vmovq_n_u8(
(motion_magnitude <= kMotionMagnitudeThreshold) ? 4 + shift_inc : 3);
const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
@ -192,10 +194,12 @@ DenoiserDecision DenoiserFilterNEON::MbDenoise(uint8_t* mc_running_avg_y,
int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
vget_low_s64(v_sum_diff_total));
int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
int sum_diff_thresh = kSumDiffThreshold;
if (increase_denoising)
if (denoise_always)
sum_diff_thresh = INT_MAX;
else if (increase_denoising)
sum_diff_thresh = kSumDiffThresholdHigh;
else
sum_diff_thresh = kSumDiffThreshold;
if (sum_diff > sum_diff_thresh) {
// Before returning to copy the block (i.e., apply no denoising),
// checK if we can still apply some (weaker) temporal filtering to

3
webrtc/modules/video_processing/util/denoiser_filter_neon.h
View File

@ -38,7 +38,8 @@ class DenoiserFilterNEON : public DenoiserFilter {
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) override;
int increase_denoising,
bool denoise_always) override;
};
} // namespace webrtc

11
webrtc/modules/video_processing/util/denoiser_filter_sse2.cc
View File

@ -139,7 +139,9 @@ DenoiserDecision DenoiserFilterSSE2::MbDenoise(uint8_t* mc_running_avg_y,
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) {
int increase_denoising,
bool denoise_always) {
unsigned int sum_diff_thresh = 0;
int shift_inc =
(increase_denoising && motion_magnitude <= kMotionMagnitudeThreshold) ? 1
: 0;
@ -211,9 +213,12 @@ DenoiserDecision DenoiserFilterSSE2::MbDenoise(uint8_t* mc_running_avg_y,
{
// Compute the sum of all pixel differences of this MB.
unsigned int abs_sum_diff = AbsSumDiff16x1(acc_diff);
unsigned int sum_diff_thresh = kSumDiffThreshold;
if (increase_denoising)
if (denoise_always)
sum_diff_thresh = INT_MAX;
else if (increase_denoising)
sum_diff_thresh = kSumDiffThresholdHigh;
else
sum_diff_thresh = kSumDiffThreshold;
if (abs_sum_diff > sum_diff_thresh) {
// Before returning to copy the block (i.e., apply no denoising),
// check if we can still apply some (weaker) temporal filtering to

3
webrtc/modules/video_processing/util/denoiser_filter_sse2.h
View File

@ -38,7 +38,8 @@ class DenoiserFilterSSE2 : public DenoiserFilter {
const uint8_t* sig,
int sig_stride,
uint8_t motion_magnitude,
int increase_denoising) override;
int increase_denoising,
bool denoise_always) override;
};
} // namespace webrtc

95
webrtc/modules/video_processing/util/noise_estimation.cc Normal file
View File

@ -0,0 +1,95 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "webrtc/modules/video_processing/util/noise_estimation.h"
namespace webrtc {
void NoiseEstimation::Init(int width, int height, CpuType cpu_type) {
int mb_cols = width >> 4;
int mb_rows = height >> 4;
consec_low_var_.reset(new uint32_t[mb_cols * mb_rows]());
width_ = width;
height_ = height;
mb_cols_ = width_ >> 4;
mb_rows_ = height_ >> 4;
cpu_type_ = cpu_type;
}
void NoiseEstimation::GetNoise(int mb_index, uint32_t var, uint32_t luma) {
consec_low_var_[mb_index]++;
num_static_block_++;
if (consec_low_var_[mb_index] >= kConsecLowVarFrame &&
(luma >> 8) < kAverageLumaMax && (luma >> 8) > kAverageLumaMin) {
// Normalized var by the average luma value, this gives more weight to
// darker blocks.
int nor_var = var / (luma >> 12);
noise_var_ +=
nor_var > kBlockSelectionVarMax ? kBlockSelectionVarMax : nor_var;
num_noisy_block_++;
}
}
void NoiseEstimation::ResetConsecLowVar(int mb_index) {
consec_low_var_[mb_index] = 0;
}
void NoiseEstimation::UpdateNoiseLevel() {
// TODO(jackychen): Tune a threshold for numb_noisy_block > T to make the
// condition more reasonable.
// No enough samples implies the motion of the camera or too many moving
// objects in the frame.
if (num_static_block_ < (0.65 * mb_cols_ * mb_rows_) || !num_noisy_block_) {
noise_var_ = 0;
noise_var_accum_ = 0;
num_static_block_ = 0;
num_noisy_block_ = 0;
#if DISPLAY
printf("Not enough samples.\n");
#endif
return;
} else {
// Normalized by the number of noisy blocks.
noise_var_ /= num_noisy_block_;
// Get the percentage of static blocks.
percent_static_block_ =
static_cast<double>(num_static_block_) / (mb_cols_ * mb_rows_);
#if DISPLAY
printf("%d %d fraction = %.3f\n", num_static_block_, mb_cols_ * mb_rows_,
percent_static_block_);
#endif
num_noisy_block_ = 0;
num_static_block_ = 0;
}
// For the first frame just update the value with current noise_var_,
// otherwise, use the averaging window.
if (noise_var_accum_ == 0) {
noise_var_accum_ = noise_var_;
} else {
noise_var_accum_ = (noise_var_accum_ * 15 + noise_var_) / 16;
}
// Reset noise_var_ for the next frame.
noise_var_ = 0;
#if DISPLAY
printf("noise_var_accum_ = %.1f, noise_var_ = %d.\n", noise_var_accum_,
noise_var_);
#endif
}
uint8_t NoiseEstimation::GetNoiseLevel() {
int noise_thr = cpu_type_ ? kNoiseThreshold : kNoiseThresholdNeon;
UpdateNoiseLevel();
if (noise_var_accum_ > noise_thr) {
return 1;
}
return 0;
}
} // namespace webrtc

56
webrtc/modules/video_processing/util/noise_estimation.h Normal file
View File

@ -0,0 +1,56 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_VIDEO_PROCESSING_UTIL_NOISE_ESTIMATION_H_
#define WEBRTC_MODULES_VIDEO_PROCESSING_UTIL_NOISE_ESTIMATION_H_
#include "webrtc/base/scoped_ptr.h"
#include "webrtc/modules/include/module_common_types.h"
#include "webrtc/modules/video_processing/include/video_processing_defines.h"
#include "webrtc/modules/video_processing/util/denoiser_filter.h"
namespace webrtc {
#define EXPERIMENTAL 0
#define DISPLAY 0
const int kNoiseThreshold = 200;
const int kNoiseThresholdNeon = 70;
const int kConsecLowVarFrame = 6;
const int kAverageLumaMin = 20;
const int kAverageLumaMax = 220;
const int kBlockSelectionVarMax = kNoiseThreshold << 1;
class NoiseEstimation {
public:
void Init(int width, int height, CpuType cpu_type);
void GetNoise(int mb_index, uint32_t var, uint32_t luma);
void ResetConsecLowVar(int mb_index);
void UpdateNoiseLevel();
// 0: low noise, 1: high noise
uint8_t GetNoiseLevel();
private:
int width_;
int height_;
int mb_rows_;
int mb_cols_;
CpuType cpu_type_;
uint32_t noise_var_;
double noise_var_accum_;
int num_noisy_block_;
int num_static_block_;
double percent_static_block_;
rtc::scoped_ptr<uint32_t[]> consec_low_var_;
};
} // namespace webrtc
#endif // WEBRTC_MODULES_VIDEO_PROCESSING_UTIL_NOISE_ESTIMATION_H_

302
webrtc/modules/video_processing/video_denoiser.cc
View File

@ -16,50 +16,144 @@ namespace webrtc {
VideoDenoiser::VideoDenoiser(bool runtime_cpu_detection)
: width_(0),
height_(0),
filter_(DenoiserFilter::Create(runtime_cpu_detection)) {}
filter_(DenoiserFilter::Create(runtime_cpu_detection, &cpu_type_)),
ne_(new NoiseEstimation()) {}
void VideoDenoiser::TrailingReduction(int mb_rows,
int mb_cols,
const uint8_t* y_src,
int stride_y,
uint8_t* y_dst) {
for (int mb_row = 1; mb_row < mb_rows - 1; ++mb_row) {
for (int mb_col = 1; mb_col < mb_cols - 1; ++mb_col) {
int mb_index = mb_row * mb_cols + mb_col;
uint8_t* mb_dst = y_dst + (mb_row << 4) * stride_y + (mb_col << 4);
const uint8_t* mb_src = y_src + (mb_row << 4) * stride_y + (mb_col << 4);
// If the number of denoised neighbors is less than a threshold,
// do NOT denoise for the block. Set different threshold for skin MB.
// The change of denoising status will not propagate.
if (metrics_[mb_index].is_skin) {
// The threshold is high (more strict) for non-skin MB where the
// trailing usually happen.
if (metrics_[mb_index].denoise &&
metrics_[mb_index + 1].denoise + metrics_[mb_index - 1].denoise +
metrics_[mb_index + mb_cols].denoise +
metrics_[mb_index - mb_cols].denoise <=
2) {
metrics_[mb_index].denoise = 0;
filter_->CopyMem16x16(mb_src, stride_y, mb_dst, stride_y);
#if EXPERIMENTAL
// Check the mb position(1: close to the center, 3: close to the border).
static int PositionCheck(int mb_row, int mb_col, int mb_rows, int mb_cols) {
if ((mb_row >= (mb_rows >> 3)) && (mb_row <= (7 * mb_rows >> 3)) &&
(mb_col >= (mb_cols >> 3)) && (mb_col <= (7 * mb_cols >> 3)))
return 1;
else if ((mb_row >= (mb_rows >> 4)) && (mb_row <= (15 * mb_rows >> 4)) &&
(mb_col >= (mb_cols >> 4)) && (mb_col <= (15 * mb_cols >> 4)))
return 2;
else
return 3;
}
static void ReduceFalseDetection(const std::unique_ptr<uint8_t[]>& d_status,
std::unique_ptr<uint8_t[]>* d_status_tmp1,
std::unique_ptr<uint8_t[]>* d_status_tmp2,
int noise_level,
int mb_rows,
int mb_cols) {
// Draft. This can be optimized. This code block is to reduce false detection
// in moving object detection.
int mb_row_min = noise_level ? mb_rows >> 3 : 1;
int mb_col_min = noise_level ? mb_cols >> 3 : 1;
int mb_row_max = noise_level ? (7 * mb_rows >> 3) : mb_rows - 2;
int mb_col_max = noise_level ? (7 * mb_cols >> 3) : mb_cols - 2;
memcpy((*d_status_tmp1).get(), d_status.get(), mb_rows * mb_cols);
// Up left.
for (int mb_row = mb_row_min; mb_row <= mb_row_max; ++mb_row) {
for (int mb_col = mb_col_min; mb_col <= mb_col_max; ++mb_col) {
(*d_status_tmp1)[mb_row * mb_cols + mb_col] |=
((*d_status_tmp1)[(mb_row - 1) * mb_cols + mb_col] |
(*d_status_tmp1)[mb_row * mb_cols + mb_col - 1]);
}
} else if (metrics_[mb_index].denoise &&
metrics_[mb_index + 1].denoise +
metrics_[mb_index - 1].denoise +
metrics_[mb_index + mb_cols + 1].denoise +
metrics_[mb_index + mb_cols - 1].denoise +
metrics_[mb_index - mb_cols + 1].denoise +
metrics_[mb_index - mb_cols - 1].denoise +
metrics_[mb_index + mb_cols].denoise +
metrics_[mb_index - mb_cols].denoise <=
7) {
filter_->CopyMem16x16(mb_src, stride_y, mb_dst, stride_y);
}
memcpy((*d_status_tmp2).get(), (*d_status_tmp1).get(), mb_rows * mb_cols);
memcpy((*d_status_tmp1).get(), d_status.get(), mb_rows * mb_cols);
// Bottom left.
for (int mb_row = mb_row_max; mb_row >= mb_row_min; --mb_row) {
for (int mb_col = mb_col_min; mb_col <= mb_col_max; ++mb_col) {
(*d_status_tmp1)[mb_row * mb_cols + mb_col] |=
((*d_status_tmp1)[(mb_row + 1) * mb_cols + mb_col] |
(*d_status_tmp1)[mb_row * mb_cols + mb_col - 1]);
(*d_status_tmp2)[mb_row * mb_cols + mb_col] &=
(*d_status_tmp1)[mb_row * mb_cols + mb_col];
}
}
memcpy((*d_status_tmp1).get(), d_status.get(), mb_rows * mb_cols);
// Up right.
for (int mb_row = mb_row_min; mb_row <= mb_row_max; ++mb_row) {
for (int mb_col = mb_col_max; mb_col >= mb_col_min; --mb_col) {
(*d_status_tmp1)[mb_row * mb_cols + mb_col] |=
((*d_status_tmp1)[(mb_row - 1) * mb_cols + mb_col] |
(*d_status_tmp1)[mb_row * mb_cols + mb_col + 1]);
(*d_status_tmp2)[mb_row * mb_cols + mb_col] &=
(*d_status_tmp1)[mb_row * mb_cols + mb_col];
}
}
memcpy((*d_status_tmp1).get(), d_status.get(), mb_rows * mb_cols);
// Bottom right.
for (int mb_row = mb_row_max; mb_row >= mb_row_min; --mb_row) {
for (int mb_col = mb_col_max; mb_col >= mb_col_min; --mb_col) {
(*d_status_tmp1)[mb_row * mb_cols + mb_col] |=
((*d_status_tmp1)[(mb_row + 1) * mb_cols + mb_col] |
(*d_status_tmp1)[mb_row * mb_cols + mb_col + 1]);
(*d_status_tmp2)[mb_row * mb_cols + mb_col] &=
(*d_status_tmp1)[mb_row * mb_cols + mb_col];
}
}
}
static bool TrailingBlock(const std::unique_ptr<uint8_t[]>& d_status,
int mb_row,
int mb_col,
int mb_rows,
int mb_cols) {
int mb_index = mb_row * mb_cols + mb_col;
if (!mb_row || !mb_col || mb_row == mb_rows - 1 || mb_col == mb_cols - 1)
return false;
return d_status[mb_index + 1] || d_status[mb_index - 1] ||
d_status[mb_index + mb_cols] || d_status[mb_index - mb_cols];
}
#endif
#if DISPLAY
void ShowRect(const std::unique_ptr<DenoiserFilter>& filter,
const std::unique_ptr<uint8_t[]>& d_status,
const std::unique_ptr<uint8_t[]>& d_status_tmp2,
const std::unique_ptr<uint8_t[]>& x_density,
const std::unique_ptr<uint8_t[]>& y_density,
const uint8_t* u_src,
const uint8_t* v_src,
uint8_t* u_dst,
uint8_t* v_dst,
int mb_rows,
int mb_cols,
int stride_u,
int stride_v) {
for (int mb_row = 0; mb_row < mb_rows; ++mb_row) {
for (int mb_col = 0; mb_col < mb_cols; ++mb_col) {
int mb_index = mb_row * mb_cols + mb_col;
const uint8_t* mb_src_u =
u_src + (mb_row << 3) * stride_u + (mb_col << 3);
const uint8_t* mb_src_v =
v_src + (mb_row << 3) * stride_v + (mb_col << 3);
uint8_t* mb_dst_u = u_dst + (mb_row << 3) * stride_u + (mb_col << 3);
uint8_t* mb_dst_v = v_dst + (mb_row << 3) * stride_v + (mb_col << 3);
uint8_t y_tmp_255[8 * 8];
memset(y_tmp_255, 200, 8 * 8);
// x_density_[mb_col] * y_density_[mb_row]
if (d_status[mb_index] == 1) {
// Paint to red.
filter->CopyMem8x8(mb_src_u, stride_u, mb_dst_u, stride_u);
filter->CopyMem8x8(y_tmp_255, 8, mb_dst_v, stride_v);
#if EXPERIMENTAL
} else if (d_status_tmp2[mb_row * mb_cols + mb_col] &&
x_density[mb_col] * y_density[mb_row]) {
#else
} else if (x_density[mb_col] * y_density[mb_row]) {
#endif
// Paint to blue.
filter->CopyMem8x8(y_tmp_255, 8, mb_dst_u, stride_u);
filter->CopyMem8x8(mb_src_v, stride_v, mb_dst_v, stride_v);
} else {
filter->CopyMem8x8(mb_src_u, stride_u, mb_dst_u, stride_u);
filter->CopyMem8x8(mb_src_v, stride_v, mb_dst_v, stride_v);
}
}
}
}
#endif
void VideoDenoiser::DenoiseFrame(const VideoFrame& frame,
VideoFrame* denoised_frame) {
VideoFrame* denoised_frame,
VideoFrame* denoised_frame_prev,
int noise_level_prev) {
int stride_y = frame.stride(kYPlane);
int stride_u = frame.stride(kUPlane);
int stride_v = frame.stride(kVPlane);
@ -71,9 +165,13 @@ void VideoDenoiser::DenoiseFrame(const VideoFrame& frame,
denoised_frame->CreateFrame(frame.buffer(kYPlane), frame.buffer(kUPlane),
frame.buffer(kVPlane), width_, height_,
stride_y, stride_u, stride_v, kVideoRotation_0);
denoised_frame_prev->CreateFrame(
frame.buffer(kYPlane), frame.buffer(kUPlane), frame.buffer(kVPlane),
width_, height_, stride_y, stride_u, stride_v, kVideoRotation_0);
// Setting time parameters to the output frame.
denoised_frame->set_timestamp(frame.timestamp());
denoised_frame->set_render_time_ms(frame.render_time_ms());
ne_->Init(width_, height_, cpu_type_);
return;
}
// For 16x16 block.
@ -81,62 +179,128 @@ void VideoDenoiser::DenoiseFrame(const VideoFrame& frame,
int mb_rows = height_ >> 4;
if (metrics_.get() == nullptr)
metrics_.reset(new DenoiseMetrics[mb_cols * mb_rows]());
if (d_status_.get() == nullptr) {
d_status_.reset(new uint8_t[mb_cols * mb_rows]());
#if EXPERIMENTAL
d_status_tmp1_.reset(new uint8_t[mb_cols * mb_rows]());
d_status_tmp2_.reset(new uint8_t[mb_cols * mb_rows]());
#endif
x_density_.reset(new uint8_t[mb_cols]());
y_density_.reset(new uint8_t[mb_rows]());
}
// Denoise on Y plane.
uint8_t* y_dst = denoised_frame->buffer(kYPlane);
uint8_t* u_dst = denoised_frame->buffer(kUPlane);
uint8_t* v_dst = denoised_frame->buffer(kVPlane);
uint8_t* y_dst_prev = denoised_frame_prev->buffer(kYPlane);
const uint8_t* y_src = frame.buffer(kYPlane);
const uint8_t* u_src = frame.buffer(kUPlane);
const uint8_t* v_src = frame.buffer(kVPlane);
uint8_t noise_level = noise_level_prev == -1 ? 0 : ne_->GetNoiseLevel();
// Temporary buffer to store denoising result.
uint8_t y_tmp[16 * 16] = {0};
memset(x_density_.get(), 0, mb_cols);
memset(y_density_.get(), 0, mb_rows);
// Loop over blocks to accumulate/extract noise level and update x/y_density
// factors for moving object detection.
for (int mb_row = 0; mb_row < mb_rows; ++mb_row) {
for (int mb_col = 0; mb_col < mb_cols; ++mb_col) {
const uint8_t* mb_src = y_src + (mb_row << 4) * stride_y + (mb_col << 4);
uint8_t* mb_dst_prev =
y_dst_prev + (mb_row << 4) * stride_y + (mb_col << 4);
int mb_index = mb_row * mb_cols + mb_col;
#if EXPERIMENTAL
int pos_factor = PositionCheck(mb_row, mb_col, mb_rows, mb_cols);
uint32_t thr_var_adp = 16 * 16 * 5 * (noise_level ? pos_factor : 1);
#else
uint32_t thr_var_adp = 16 * 16 * 5;
#endif
int brightness = 0;
for (int i = 0; i < 16; ++i) {
for (int j = 0; j < 16; ++j) {
brightness += mb_src[i * stride_y + j];
}
}
// Get the denoised block.
filter_->MbDenoise(mb_dst_prev, stride_y, y_tmp, 16, mb_src, stride_y, 0,
1, true);
// The variance is based on the denoised blocks in time T and T-1.
metrics_[mb_index].var = filter_->Variance16x8(
mb_dst_prev, stride_y, y_tmp, 16, &metrics_[mb_index].sad);
if (metrics_[mb_index].var > thr_var_adp) {
ne_->ResetConsecLowVar(mb_index);
d_status_[mb_index] = 1;
#if EXPERIMENTAL
if (noise_level == 0 || pos_factor < 3) {
x_density_[mb_col] += 1;
y_density_[mb_row] += 1;
}
#else
x_density_[mb_col] += 1;
y_density_[mb_row] += 1;
#endif
} else {
uint32_t sse_t = 0;
// The variance is based on the src blocks in time T and denoised block
// in time T-1.
uint32_t noise_var = filter_->Variance16x8(mb_dst_prev, stride_y,
mb_src, stride_y, &sse_t);
ne_->GetNoise(mb_index, noise_var, brightness);
d_status_[mb_index] = 0;
}
// Track denoised frame.
filter_->CopyMem16x16(y_tmp, 16, mb_dst_prev, stride_y);
}
}
#if EXPERIMENTAL
ReduceFalseDetection(d_status_, &d_status_tmp1_, &d_status_tmp2_, noise_level,
mb_rows, mb_cols);
#endif
// Denoise each MB based on the results of moving objects detection.
for (int mb_row = 0; mb_row < mb_rows; ++mb_row) {
for (int mb_col = 0; mb_col < mb_cols; ++mb_col) {
const uint8_t* mb_src = y_src + (mb_row << 4) * stride_y + (mb_col << 4);
uint8_t* mb_dst = y_dst + (mb_row << 4) * stride_y + (mb_col << 4);
int mb_index = mb_row * mb_cols + mb_col;
// Denoise each MB at the very start and save the result to a temporary
// buffer.
if (filter_->MbDenoise(mb_dst, stride_y, y_tmp, 16, mb_src, stride_y, 0,
1) == FILTER_BLOCK) {
uint32_t thr_var = 0;
// Save var and sad to the buffer.
metrics_[mb_index].var = filter_->Variance16x8(
mb_dst, stride_y, y_tmp, 16, &metrics_[mb_index].sad);
// Get skin map.
metrics_[mb_index].is_skin = MbHasSkinColor(
y_src, u_src, v_src, stride_y, stride_u, stride_v, mb_row, mb_col);
// Variance threshold for skin/non-skin MB is different.
// Skin MB use a small threshold to reduce blockiness.
thr_var = metrics_[mb_index].is_skin ? 128 : 12 * 128;
if (metrics_[mb_index].var > thr_var) {
metrics_[mb_index].denoise = 0;
// Use the source MB.
filter_->CopyMem16x16(mb_src, stride_y, mb_dst, stride_y);
} else {
metrics_[mb_index].denoise = 1;
// Use the denoised MB.
filter_->CopyMem16x16(y_tmp, 16, mb_dst, stride_y);
}
} else {
metrics_[mb_index].denoise = 0;
filter_->CopyMem16x16(mb_src, stride_y, mb_dst, stride_y);
}
// Copy source U/V plane.
const uint8_t* mb_src_u =
u_src + (mb_row << 3) * stride_u + (mb_col << 3);
const uint8_t* mb_src_v =
v_src + (mb_row << 3) * stride_v + (mb_col << 3);
uint8_t* mb_dst_u = u_dst + (mb_row << 3) * stride_u + (mb_col << 3);
uint8_t* mb_dst_v = v_dst + (mb_row << 3) * stride_v + (mb_col << 3);
#if EXPERIMENTAL
if ((!d_status_tmp2_[mb_row * mb_cols + mb_col] ||
x_density_[mb_col] * y_density_[mb_row] == 0) &&
!TrailingBlock(d_status_, mb_row, mb_col, mb_rows, mb_cols)) {
#else
if (x_density_[mb_col] * y_density_[mb_row] == 0) {
#endif
if (filter_->MbDenoise(mb_dst, stride_y, y_tmp, 16, mb_src, stride_y, 0,
noise_level, false) == FILTER_BLOCK) {
filter_->CopyMem16x16(y_tmp, 16, mb_dst, stride_y);
} else {
// Copy y source.
filter_->CopyMem16x16(mb_src, stride_y, mb_dst, stride_y);
}
} else {
// Copy y source.
filter_->CopyMem16x16(mb_src, stride_y, mb_dst, stride_y);
}
filter_->CopyMem8x8(mb_src_u, stride_u, mb_dst_u, stride_u);
filter_->CopyMem8x8(mb_src_v, stride_v, mb_dst_v, stride_v);
}
}
// Second round.
// This is to reduce the trailing artifact and blockiness by referring
// neighbors' denoising status.
TrailingReduction(mb_rows, mb_cols, y_src, stride_y, y_dst);
#if DISPLAY // Rectangle diagnostics
// Show rectangular region
ShowRect(filter_, d_status_, d_status_tmp2_, x_density_, y_density_, u_src,
v_src, u_dst, v_dst, mb_rows, mb_cols, stride_u, stride_v);
#endif
// Setting time parameters to the output frame.
denoised_frame->set_timestamp(frame.timestamp());

20
webrtc/modules/video_processing/video_denoiser.h
View File

@ -14,6 +14,7 @@
#include <memory>
#include "webrtc/modules/video_processing/util/denoiser_filter.h"
#include "webrtc/modules/video_processing/util/noise_estimation.h"
#include "webrtc/modules/video_processing/util/skin_detection.h"
namespace webrtc {
@ -21,18 +22,25 @@ namespace webrtc {
class VideoDenoiser {
public:
explicit VideoDenoiser(bool runtime_cpu_detection);
void DenoiseFrame(const VideoFrame& frame, VideoFrame* denoised_frame);
void DenoiseFrame(const VideoFrame& frame,
VideoFrame* denoised_frame,
VideoFrame* denoised_frame_track,
int noise_level_prev);
private:
void TrailingReduction(int mb_rows,
int mb_cols,
const uint8_t* y_src,
int stride_y,
uint8_t* y_dst);
int width_;
int height_;
CpuType cpu_type_;
std::unique_ptr<DenoiseMetrics[]> metrics_;
std::unique_ptr<DenoiserFilter> filter_;
std::unique_ptr<NoiseEstimation> ne_;
std::unique_ptr<uint8_t[]> d_status_;
#if EXPERIMENTAL
std::unique_ptr<uint8_t[]> d_status_tmp1_;
std::unique_ptr<uint8_t[]> d_status_tmp2_;
#endif
std::unique_ptr<uint8_t[]> x_density_;
std::unique_ptr<uint8_t[]> y_density_;
};
} // namespace webrtc

2
webrtc/modules/video_processing/video_processing.gypi
View File

@ -40,6 +40,8 @@
'util/denoiser_filter.h',
'util/denoiser_filter_c.cc',
'util/denoiser_filter_c.h',
'util/noise_estimation.cc',
'util/noise_estimation.h',
'util/skin_detection.cc',
'util/skin_detection.h',
],